Key Points
- Behaviour often follows geometry, not just composition - AI can sweep forms before lab work.
- Use AI results as time-shifted hypotheses: unbuildable today, viable as biotech/process advances arrive.
- Bake compliance in: link prompts to LCA/DPP constraints so options start low-impact.
- LLMs can map local materials and regulations into workable taxonomies and specs.
Full interview
1. Your paper maps AI’s role in conceptual workflows. How might generative systems inform or provoke new material imaginaries before any empirical research begins?
Generative AI tools can help uncover prior material research, suggesting novel combinations or mixes worth exploring. While materials themselves matter, behaviour often depends on geometry. AI algorithms can explore vast design spaces of geometries and simulate how materials might act before any real-world testing.
2. As a researcher, how do you evaluate the epistemic value of AI-generated outputs when they suggest materials that do not yet exist or whose performance is unknown?
Novel ideas always hold value. To assess whether they could work in reality, researchers rely on their applied and theoretical knowledge. Another question is when a material could work. For instance, mycelium, a living material widely used in architecture, is prone to mould and must be baked to stabilise it. Future advances in genetic engineering could allow for strains that stop growing naturally, as already achieved with tomatoes. AI might propose concepts that are unfeasible today but viable tomorrow, much like discoveries in mathematics that only reveal their applications decades later.
3. In exploring layered languages: textual, architectural, and code. What potential do you see for AI models to generate not just speculative forms but early indicators of material function or behaviour?
AI language models act as repositories of knowledge and can help map local materials available on site. Other AI models can predict how these materials interact with environmental and regulatory factors such as climate, radiation, or seismic requirements. Trained properly, they could act as design assistants, proposing compliant material solutions from the start.
4. The use of text-to-text models allowed for semantic expansion. How might this linguistic layer assist in constructing future taxonomies for emerging materials?
Text-to-text models could contribute at every stage. They can help craft design briefs, retrieve previous research, simulate material behaviour, and even communicate with digital fabrication systems. The linguistic layer acts as a bridge between thought and production.
5. How can AI models trained on architectural datasets be tuned or expanded to include data on bio-materials, circular components, or energy-sensitive composites?
Architectural datasets should integrate detailed characterisation of bio-materials. On a lifecycle level, the upcoming Digital Product Passports in the EU could offer valuable data on material reuse and degradation. While the built environment still lacks consistent data, these initiatives can help AI models link material selection to environmental impact.
6. What are the methodological implications of treating AI as a speculative co-author in material research, particularly when outcomes veer into the unreal or non-fabricable?
Architecture inherently deals with the “not yet”. Our role is to imagine futures, whether immediate or distant. Before we can construct responsible worlds, we must first imagine them. AI supports this by testing future possibilities, such as off-planet habitats or bioengineered matter, before these ideas become technically feasible.
7. You discuss language as a bridge. How do annotations in AI training corpora influence how materials are represented or prioritised in generative outputs?
Annotations are often produced in annotation farms with limited oversight, which can introduce bias or imprecision. Perfect datasets don’t exist. Yet, language remains a powerful connector between code and fabrication. We might soon see AI models communicating directly with machines such as laser cutters or 3D printers, which will redefine what can be expressed through language in design.
8. In conceptual competitions like eVolo, where tangibility is deferred, how do you critically assess the plausibility or rigour of material propositions born from AI?
Competitions such as eVolo are more about speculative futures than construction feasibility. They operate like architectural science fiction, offering spaces for experimentation. As Bruno Latour suggested, our era faces a crisis of imagination, where we struggle to define the futures we want and instead retreat to nostalgia.
9. Looking at image-to-image workflows, what limitations did you observe in current models’ capacity to encode or infer materiality, texture, or surface logic?
The limitations are significant. Image-to-image models generate visual representations, not material knowledge. For architecture, we need domain-specific models that encode material characterisation such as density, porosity, and surface behaviour. The goal is to move from images to inference.
10. How might researchers better integrate environmental metrics, like embodied carbon or life cycle assessments, into the training sets of generative tools to enhance material relevance?
Environmental metrics should serve as constraints in generative systems. By embedding embodied carbon and LCA data, AI can guide designers toward lower-impact outcomes. Yet these metrics are only as strong as the data behind them. For new materials, models may initially act as hypotheses that evolve alongside empirical research.
11. In your view, should generative models remain speculative provocateurs, or can they evolve into viable forecasting tools within material innovation pipelines?
Both roles are necessary. Speculation inspires, while forecasting refines. Generative models should oscillate between the two, using imagination to guide practical innovation.
12. What role do you see for interdisciplinary collaboration, particularly with materials scientists and computational engineers, in shaping the next generation of AI-augmented architectural research?
Inter- and trans-disciplinary collaboration will be essential. Architecture has always integrated diverse expertise. Future design teams will include computer scientists, materials scientists, biotechnologists, and agricultural researchers. Architects will lead these coalitions, balancing multiple ways of knowing and building coherent frameworks for innovation.
